When playing with basic dielectric substrates, I found that the model computed with Nvidia, indeed, requires less memory. I did not find serious troubles with the solution, except Nvidia RTX 5000 which I used is blazingly fast.

...And yes, AFAIK RTX cards are better in mixed precision than GTX, because they can use different cores (different precision) simultaneously, which older GTX cards can not. Or how it is explained in gaming industry. I am not very familiar about hardware details.